Abstract

Despite ongoing progress in deep learning for fault diagnosis, several obstacles hinder its effective implementation in the industry. Firstly, obtaining labeled data is challenging due to the high costs and time required for annotation, leading to a scarcity of labeled datasets. Secondly, even when labeled data is available, there can be significant distribution shift because motors in real-world applications operate under diverse torque conditions. To address these issues, this paper introduces a new self-supervised feature learning approach called notch filter augmentation-based self-supervised learning. This approach involves a self-supervised learning framework with the augmentation using a notch-filter that can be applied to stator current signals. It is designed to extract fault features under various torque conditions by leveraging knowledge in the frequency domain and transferring it to the time domain. This enables the model to learn fault features from the signals, leading to robust fault diagnosis under various torque conditions. The effectiveness of the proposed method is validated through its application to a permanent magnet synchronous motor (PMSM) dataset under two different torque conditions, demonstrating its ability to accurately identify fault features. Validation results indicate that the proposed method performance exceeds existing self-supervised and supervised models under operating conditions of varying torque and limited availability of labeled of labeled data.